Frelzu

Human curiosity. Artificial intelligence. A front-row seat to what's next.

[Frelzu @Frelzu]

Not charts after the match. Real-time decisions.

How AI started calling the shots in F1, football, tennis and the NFL. https://t.co/J8ZxTW8GrR

[Frelzu @Frelzu]

@Plutkar That's the real test. 👀

200T tokens prove people were willing to use it. But long-running agents need more than cheap inference — they need consistency across hundreds of steps.
If MiniMax can deliver both, that's where things get interesting.

[Frelzu @Frelzu]

200 TRILLION TOKENS.

ONE MYSTERY AI MODEL.

Now we know who built it.

“Space Bunny” quietly processed 200T+ tokens while running anonymously.

Today, MiniMax confirmed it was an early version of M3.1 Flash.

And the economics are aggressive:

→ $0.15 / 1M input tokens
→ $0.60 / 1M output tokens
→ 70% lower agent-task costs, according to MiniMax
→ Nearly 3x faster inference
Here's the bigger story:
AI labs aren't just competing to build smarter models.
They're competing to make intelligence cheaper to deploy.

And that could put serious pressure on premium AI pricing.

The next AI war is about COST PER TASK.

Not just benchmark scores.

[MiniMax (official) @MiniMax_AI]

MiniMax M3.1 Flash: Space Performance + Bunny Speed

🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.

💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.

API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.

Model weights will arrive on Oct 23.

API: https://t.co/fHRdSV73Hr
M Plan: https://t.co/XfJWSyauda
MiniMax Code for desktop: https://t.co/SGmLTI2HQm

[Frelzu @Frelzu]

@Amos_FACT

[Frelzu @Frelzu]

@starmexxx AI subscriptions are becoming the new SaaS bloat.
The winners won't be the people paying for 10 different models.

They'll be the ones making 3 models work together.

[Frelzu @Frelzu]

@0xCortiq Not at ATB ) This one was raised on 200 trillion tokens, not carrots.

[Frelzu @Frelzu]

@rost_roma That's the comparison worth watching. 
Same prompt, three models — but the real winner should be measured by quality, completion time, and total cost.
Looking forward to your results!

[Frelzu @Frelzu]

@Amos_FACT hello, follow me I'll follow back

[Frelzu @Frelzu]

@nar96265 hey everyone

[Frelzu @Frelzu]

@Shileisaboy Hi

[Frelzu @Frelzu]

@hexobuilds Which of the 10 mistakes costs the most in practice: bloated context, no spec up front, or skipping review of the output?

[Frelzu @Frelzu]

@Flinzu An old Unity sim I sunk way too many hours into. Always wondered how the economy actually worked under the hood. Curious how much of the logic survives the read.

[Frelzu @Frelzu]

@gippp69 The best part isn’t the bots, it’s the order: fewest bots first, read-only first, approval before anything irreversible. Most people do it backwards and give the agent keys on day one.

[Frelzu @Frelzu]

Why does every second brain die after about three months?

Not because you stop caring. Because keeping it current is a full-time job.

The fix: stop filing, start compiling.

raw/ → wiki/ → output/, with CLAUDE.md in the center. Claude reads each source once, links it to what you already have, flags contradictions, and writes you a brief.

One catch, straight from the article: it only gets good after around 50 sources. Before that it’s just a search engine. Don’t quit early.

[rvaniaaa @rvaniaaaa]

https://t.co/wBLzP25PUm

[Frelzu @Frelzu]

@buswe_com Right, set it once and the map is always there. Good addition, thanks.

[Frelzu @Frelzu]

Stop building your second brain in Notion and Obsidian.

The version that actually works with AI is 3 folders and 1 text file.

Here’s the idea:

Give Claude Code a folder of plain .md files and one index file that says where everything lives. It reads the map first, then goes straight to the right note. The more you add, the smarter it gets, like compound interest for your knowledge.

The setup:

→ Create a folder and open it in Claude Code
→ Add three subfolders: notes, people, projects
→ Put a CLAUDE.md in the root as the map
→ One topic = one file, named after what’s inside
→ Drop in what you learn, ask questions anytime
No databases. No bloated apps. Just text files.

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Frelzu @Frelzu]

@Kaggii_ I’m only getting started with this kind of setup, so I can’t say yet. Your post made me realize that’s something I should build in from the start.

[Frelzu @Frelzu]

Sounds interesting, but I’m still skeptical. “3,205 tests passed and all 63 compiled bytes matched” is a pretty narrow case. On a real closed-source product with proper obfuscation, anti-debug, and a bunch of dependencies this usually falls apart quickly.

Also curious about the practical side — how many tokens and how much time does it actually take for a more complex function? And how often do you end up manually fixing what the model hallucinated in the call graph or data structures?

If anyone’s already tried this on actual proprietary binaries (not toy examples), how close does it really get you to “rebuild without source” versus just “understand roughly what’s going on”?

[Frelzu @Frelzu]

@0xCodila How does it handle conflicts when two sub-agents return contradicting results? That’s usually where orchestrators break.